utils

Tensor utilities: clustering, voxelization, geometry, serialization, metrics, and I/O.

Modules:

• box3d –

  Generic 3D oriented-box geometry shared by detection models and their evaluation.

• cluster –

  Neighbor search and grouping: kNN, FPS, radius queries, and local grids.

• conversion –

  Conversions between iterables, batch index representations (offset, bincount, cu_seqlens), and tensor formats.

• data –

  Data loading: standard sample keys, packed-batch collation, and the point cloud data loader.

• diffusion –

  Denoising diffusion schedules and samplers.

• ensemble –

  Prediction reducers for ensemble / TTA / voting workflows.

• geometry –

  Geometric operations: bounding boxes, rotations, point transforms, vertex normals, and spherical sampling.

• heatmap –

  Gaussian center-heatmap targets for center-based 3D detection heads.

• hilbert –

  Hilbert curve encoding and decoding between coordinates and curve indices.

• imports –

  Optional dependency detection, lazy imports, and availability flags.

• io –

  File reading and writing for point cloud data: JSON, OFF meshes, and safetensors.

• metrics –

  Segmentation, detection, and instance metrics: IoU, accuracy, and average precision variants.

• misc –

  Miscellaneous helpers for parallel mapping and nested attribute access.

• neighbors –

  Gaussian kernel density estimation over packed batches.

• octree –

  Octree construction, interpolation, and upsampling helpers built on ocnn.

• ops –

  Tensor operations on packed batches: safe division, softmax, voxel hashing, interpolation, and decimation.

• optim –

  Optimizer utilities: parameter-group construction.

• random –

  Random seeding and determinism control.

• serialization –

  Space-filling curve serialization of voxel coordinates using Z-order and Hilbert encodings.

• state_dict –

  Checkpoint reading, adaptive state_dict loading, and key-mapping transforms.

• types –

  Shared type aliases, TypedDict definitions, and enums used across the package.

• voxelization –

  Dense and sparse voxelization with trilinear devoxelization for packed point clouds.